The distilled accurate tier

RecursiveAI.OneNlp.Models.Multilingual: one six-layer, width-384 student of mmBERT-base (35.5 M weights, distilled in PyTorch from mmBERT-base and exported once; recipes/accurate/student-v1), fine-tuned with per-language heads for English, German, French, Spanish, Portuguese, Italian, Dutch, Turkish and Chinese (tagger, morphologizer, edit-tree lemmatizer, biaffine parser, entity recognizer), plus an open-type entity head on the frozen student. The frozen student is also published alone (RecursiveAI.OneNlp.Models.Multilingual.FrozenEncoder, FrozenStudentEncoder.Create()).

• Recipe: recipes/accurate/student-multilingual (train, with --export DIR writing the encoder and one head file per language; export-open re-saves the open-type head over the frozen encoder). The languages are data: --langs names codes, --languages tools/fetch/ud/languages.tsv supplies the gating treebank of any other language whose data has been fetched, so adding a language is a line in that file and a fetch.
• Fine-tuned for the heads, frozen for the open-type head. On English-EWT UPOS a fine-tuned student scored 95.88 and a frozen one 93.01 (recipes/accurate/student-v1/RESULTS.md); the open-type head was trained, and checked, on the frozen student and keeps it.
• Training run: seed 1, 2,000 steps of about 3,000 words over the nine languages (language-balanced), learning rate 2e-4 chosen on development against 1e-4 (80.40 against 77.85 mean development score at step 1,250); the development score was still rising at the end, so the model is under-trained, not converged. About 2 hours on four cores. Provenance: the package's PROVENANCE.md.
• Size and speed (indicative, one thread, a loaded 16-core x64 machine, load about 9): the packages total 482 MB (main 172, Encoder 131, FrozenEncoder 136, Segmenters 43). Full pipeline (tokenize, tag, lemmatize, parse, entities): 1,153 tokens/s on English and 1,162 on German test sentences (400 sentences each, best of three), peak RSS about 0.8 GB. The first Nlp.Analyze of a short sentence took 37 ms and the process stayed at 47 MB (the weights are mapped from the assembly, so the pages a sentence does not touch are not read yet).
• Training cost against PyTorch (indicative): one fine-tuning step of the student, 24 sentences of 32 pieces, 8 threads, same machine: 0.41 s in C# (forward 0.11, backward 0.24, clipping 0.03, AdamW 0.04) against 0.13 s in PyTorch 2.11 on CPU, 3.2 times. Gradient clipping and AdamW account for a disproportionate part (about 0.07 s against 0.01 s); both are on the OneModel backlog.

Results

Every row is information: the base-size tier is the one that gates. Per language and input mode, with the fast tier's rows and the published Stanza, UDPipe 2 and Trankit numbers beside ours: distilled-tier-rows.md. Test splits were read once for this model; development sets chose the learning rate and the checkpoint.

Entities (micro F1, test, gold tokens, bootstrap 95% interval over sentences, 1,000 resamples, seed 20261106)

Entity test sets are capped at 3,000 sentences sampled evenly, so rows are not the fast tier's full-test rows.

Where it stands

• It is ahead of the fast tier on UPOS in English raw text, Dutch and Turkish and on entities in English, German, Portuguese, Turkish and Chinese; it trails it on lemmas everywhere (the lemmatizer here has less capacity than the fast tier's), on parsing (LAS 4 to 15 points), on French, Spanish, Italian and Dutch entities, and on Chinese tagging and parsing (UPOS 85.56 against 90.41, LAS 54.05 against 68.76).
• It trails the published UDPipe 2 and Stanza numbers by a wide margin on every row: a short run of a six-layer student. The base-size tier (accurate-tier.md) is the model for accuracy.
• Sentence splitting comes from the parser and is weakest in English (65.4 on EWT raw text) and Chinese.
• Hindi has no allowed training data and is not in this package.